Correcting measurement error bias in interaction models with small samples